Evidence Model and Rules

Evidence labels

LabelMeaningExample
measuredread or timed on this host nowCPU, RAM, disk, CPU-loop timing
heuristiccomputed from stated assumptionsmodel fit
runtime_reporteda local runtime said soOllama eval_count
not_runnot done; never a passlive inference, egress test
unknowncannot be determinedAMD GPU, NPU
blocked / failedrefused by policy / attempted and failedremote endpoint, timeout

Audit evidence tiers

TierProvesDoes not prove
MEASURED_LOCALvalues this host reported at that timeother hardware; that a model will run
HEURISTICa stated formula's outputload success, speed, quality
MOCK_FIXTUREprobe logic against a local mock servicebehaviour of a real Ollama or llama.cpp
RULE_CHECKendpoint rules and the data inventorythat a runtime sends no traffic

The seal

evidence.sha256 covers the canonical payload. It detects change. It is not a signature, does not name an author and can be recomputed by someone who edits the report. Signed attestations are planned.

The ten operating rules (rules/)

RuleEnforced by tests?
read-only-by-defaultyes: no writes except explicit outputs
no-cloud-fallbackyes: no code path reaches a non-loopback host
loopback-boundaryyes: A01, A02, unit tests
no-fabricated-gpuyes: D01-D03, fuzz
benchmark-honestyyes: F01, F02
model-fit-uncertaintyyes: C01, C02
secret-and-prompt-safetyyes: B02, E03
bounded-executionyes: limits tests
provenance-and-evidencepartly: seal and labels tested; signatures planned
human-release-approvalinstruction-only (process rule)

Comparability

Two benchmark reports are compared only when workload, iteration count, platform and architecture match; otherwise the tool refuses. Different models, quantizations, context lengths or hardware are never ranked.

Claims discipline

A statement appears in docs only if a test, a proof file or a CI run backs it. Everything else is marked planned, unknown or not run.